The accuracy of K-means algorithm is greatly influenced by initial clustering center which is related to the density distribution of the sample to some extent. In order to obtain the gradually optimized initial clustering center, the present study implement the loop fusion of sub regions based on their density and correlation by dividing the region into sub ones. The improved K-means algorithm is tested in two dimensions, and the results show that the improved algorithm obtains more optimized initial clustering center with higher accuracy of clustering.